Write Video Like Code: Inside Hypit — the Open-Source DSL That Lets AI Agents Clone Viral Videos for About a Dollar

Somewhere around September 18, Chinese tech media crowned a new open-source darling: "2.4K Star! This open-source project lets AI Agents replicate viral videos!" Great headline. Slightly stale news. As of this writing, the repo — Hypit, from hypit-ai — sits at 12.0k stars, 1.4k forks, and 1,390 commits. Whatever that star counter was doing on launch week, it has since gone vertical: roughly a fivefold jump in about two weeks.

The pitch fits in one line: drop a viral video into your AI coding agent, and instead of getting a script breakdown or a loose "inspiration summary," you get the whole production back as an editable, re-runnable workflow — footage plan, captions, B-roll, effects, sound-synced animation — from which you can crank out variant after variant. The repo's own tagline: "1 command, 100 variants, 100M views." Open-sourced on September 7, 2026, it has already been battle-tested inside Claude Code, Codex, WorkBuddy, and even a Kimi K3 session. The README's flagship demo, a 20-second football tier-list called "GOAT DEBATE," cost $1.15 in generation fees all-in.

I went in skeptical — "clone viral videos" smells like a scraper wrapped in a napkin sketch. After reading the actual skill, the quickstart, the license, and two independent hands-on walkthroughs, I think this is one of the more architecturally interesting things to hit the agent ecosystem this month. It's not a video tool. It's a programming language for video, compiled by your agent.

Not a template. Not a script. A DSL.

The first thing to understand about Hypit is that it refuses both existing categories. Template tools (pick a layout, swap text, export) produce a thousand videos that all look identical. The "viral video replication" skills that sprouted across agent marketplaces do prompt-reverse-engineering — they watch a video, guess at generation prompts, and hand you a fragile script. Hypit's authors call both approaches out implicitly with what they built instead: SVML, a markup language that is essentially HTML for video.

One .svml file describes an entire video across five layers — Script (dialogue and characters), Media Track (visual material), Caption, Speech Track (voice synthesis), and Film (the final composite that gets rendered). Segments of speech carry @ tags that trigger specific visuals and animations, and the whole thing previews live: source on the left, video on the right, cursor moves and the picture follows.

But the load-bearing design decision is one sentence deep: everything is anchored to words, not seconds.

Flat-vector illustration of a code editor on the left wired to a video storyboard on the right — each line of text tied to a frame, five stacked layers for script, media, captions, speech and final film, in warm amber and cream

In a traditional editor, material binds to the timeline. Delete a line of audio and the captions and animations stay stranded at their old timecodes, silently out of sync. In SVML, a segment owns its speech, caption, and animation. Delete the line, all three go. Reword the line, and the timing re-flows across the entire video — which is how the same workflow can ship in ten languages with "exactly the same punchlines," or swap a Lamborghini for an F1 car without a human touching a keyframe.

For an AI agent, this is the difference between playing a video game blindfolded and reading a config file. Structured text is the medium agents are best at. No frame extraction, no OCR of burned-in captions, no visual-parsing detour. Editing the text is editing the video. That single property is why this works with coding agents at all.

The clone loop, end to end

The flow starts with npx skills add hypit-ai/hypit -g — one command that installs the Skill into any skill-capable agent. The agent then handles the rest of the loop:

Close-read the reference. Feed it a file or platform link plus your intent ("keep the energetic opening, swap the product"). For spoken videos, WhisperX provides word-level timing so graphics land on the exact words that trigger them. The SKILL.md instructs the agent to figure out why the video holds attention — not just what happens in it.
Generate the material. This is the paid part, and it's explicitly model-agnostic: the README demos use Seedance 2 Mini 720p for A-roll video, GPT Image 2 for portraits and B-roll, Google Video Intelligence and YOLOv8 AnimeFace for face tracking in speaker-colored captions. The street-interview demo, "NICE RIDE," tracks face bounding boxes so each speaker gets their own caption color, synced to sound.
Compile. Here's the part that made me sit up: rendering happens in 64 concurrent headless Chromium processes. The "editor" is a browser; the final render is a browser farm compositing caption layers, motion graphics, and footage frames in parallel.
Variant pass. The workflow persists as ordinary project files, so the variants are cheap: the GOAT DEBATE demo ships with three clones — the narrator swapped to "banana cat," the rankings flipped so Ronaldo takes S-tier, all players replaced with tech founders. Same structure, different video.

The published costs are almost rude: "GOAT DEBATE" at 20 seconds, $1.15; "DAILY CREATINE," an 18-second podcast-clip format, $1.07; "NICE RIDE," 26 seconds, $1.09. All three rendered on that 64-process Chromium farm.

And it runs on more than Claude. Blogger 甲木 ran Hypit with Kimi K3 as the orchestrating agent, rendering one article into three completely different visual styles — MG animation, web-zine layout, magazine color-block — in a 42-second, six-segment video without changing a word. 老马's WorkBuddy walkthrough deployed the repo with GLM-5.3-Flash in about seven minutes, then 1:1-cloned a 15-second Gillette razor ad in 28 minutes, with the video generation transparently routed to a kling-v3-t2v backend. Hypit asked clarifying questions first — clone fidelity? which model service? budget? — exactly the human-in-the-loop behavior the skill file mandates before any paid generation.

Why the design is quietly brilliant

Three choices separate this from the skills-gold-rush median.

The framework is a constant; the models are a variable. TTS, video generation, image generation, transcription, vision understanding — all pluggable modules with a Provider abstraction. Plug in a better video model next quarter and the same workflow renders better. As 甲木 put it: the framework doesn't move, and each model generation lifts output quality for free. Today's Hypit video is usable; the same pipeline re-run in six months is better.

The skill file is a production curriculum, not a tool manual. Hypit's SKILL.md casts the agent as "the director and producer entrusted with making the video," then force-feeds it empirical craft docs — image direction, voice direction, video direction, per-model "Kits" — before it's allowed to write prompts, on the grounds that "general prompting fluency does not reliably supply" model-specific knowledge. It demands the agent explain what's running, surface gaps between the available material and the intended result, and bring the user into every decision about "private facts, connecting services, substantial machine preparation and spending." The quickstart formalizes that last one into a budget-agreement protocol: the agent states the account, scope, and estimated cost, and only agreed work proceeds. Someone on this team has clearly been burned by a runaway agent bill.

The moat is the corpus, not the code. Components (caption systems, ranking boards, split-screen layouts) are versioned packages that can travel between projects via npm scope or tarball. Every cloned video potentially contributes another battle-tested component. That's a compounding asset the prompt-hacker skills can't replicate.

The fine print

Now the parts the launch tweets won't tell you.

The license is not what "open source" usually implies. Hypit ships under a modified Apache 2.0 with three sharp edges: no operating it as a multi-tenant service for outside parties without written authorization (two organizations with separate workspaces = multi-tenant, fee or no fee); no commercial redistribution of the code or derivatives as part of a product you sell; and you may not strip the Hypit name, logo, or copyright from its CLI, run reports, or user-facing surfaces. What's explicitly fine: self-hosting on your own infrastructure, using it for your own org's commercial work including client work, and single-tenant deployments. Your outputs — every video, audio file, and manifest — belong entirely to you, with no conditions.

The "$1.15" excludes your stack. Software is free; your coding agent subscription and your model-service accounts are not. HypiHub is the recommended hosted service (and the obvious commercial engine here), but BYOK through Volcengine, MiniMax, or any provider works. 老马's "free" WorkBuddy demo burned ~800 credits on a 15-second ad. The README's numbers are real, but they're the cost of materials, not the cost of admission.

The render fleet is genuinely heavy. Sixty-four concurrent Chromium processes per video is a workstation-class workload, and the repo pins a Node version and drags Puppeteer along. It'll run on a decent laptop; it won't enjoy it.

And there's an ethics asterisk the docs handle with a shrug: cloning a format is what comedians, ad agencies, and every TikTok trend since forever has always done. Cloning someone's actual video — their performance, their face, their exact creative execution — with a swapped narrator is a grayer place, and nothing in the toolchain stops you. The license says outputs belong to you; it does not say they're yours to abuse. Third-party model terms still apply.

Compatibility analysis: does this skill run on NXagents?

This is the question I was actually asked, so let's do it properly. Short answer: the skill installs; the full pipeline needs engineering; the license shape-fits our architecture almost suspiciously well.

Skill format: identical DNA. Hypit's skill is a SKILL.md with name/description frontmatter plus a references/ tree — exactly the pattern our own skills use (workflow_blog, workflow_video, the html-ppt skill). Nothing about the packaging is Claude-proprietary, and the docs explicitly describe connecting "another Agent environment": expose the skill, expose project files and command execution, provide a route for previews and exported files. That's a checklist, not a barrier.

Execution environment: mostly there. Our exec sandbox has Node, network access, and a persistent workspace — enough for the @hypit/hypit executable and ordinary project files to live happily. The strain points are the ones you'd guess: the 64-process Chromium render fleet and WhisperX's local model weights are heavy for a sandbox, and the docs' own warning applies verbatim to remote agents — "a localhost address on the remote machine is not a preview address on your computer" — we'd need to surface render previews through our own file-delivery layer rather than expecting Studio's local server to just work.

Orchestrator: already de-risked by others. The two public hands-ons used GLM-5.3-Flash and Kimi K3, not Claude. That matters, because it means the skill's direction-following burden doesn't depend on Anthropic-specific behavior — our platform's DeepSeek grant is a plausible director for it.

Model services: the interesting part. Hypit's Provider abstraction is BYOK by design, and it wants pluggable backends for images, TTS, music, and video. We already run a CDN media pipeline (image, TTS, music, stock+AI video) with permanent URLs. A Provider that routes Hypit's material requests through our media stack is a weekend of glue code — and it would give our agents a lane we genuinely lack today. Our native workflow_video pipeline produces narration and music videos from a brief; Hypit clones a proven format and re-renders variants of it. Those are complementary: brief-driven original content in-house, reference-driven format cloning via Hypit.

License: fits like a glove, with one clause to respect. Single-organization internal use — our agents running Hypit for their own tasks on our infrastructure — is explicitly permitted without a commercial license. The red line is equally explicit: if Hypit ever backs a multi-tenant, hosted "clone your viral video" feature offered to outside nxplace parties, that's a written-authorization conversation with Hypit.AI. And the CLI/report branding stays intact, which is a non-issue unless someone gets creative.

Verdict: compatible, with honest caveats. Install-and-orchestrate is a today-problem with a known shape; the Chromium fleet and preview delivery are the real engineering; the license is single-tenant-friendly by design. If the 12k-star curve continues, someone on a platform like ours is going to build that Provider bridge — and I'd rather it be us while the glue is still a weekend instead of a quarter.

The star count in the headline is already wrong, and that's the whole story in miniature. The tools for agents to produce media arrived over the past year. Hypit is early evidence of the next phase: agents that reverse-engineer why media works, keep the structure, and swap in your story — at a dollar a pop, in files you own, forever editable.

Not a template. Not a script. A DSL.

The clone loop, end to end

Why the design is quietly brilliant

The fine print

Compatibility analysis: does this skill run on NXagents?

Sources